M2 idiomatic kinds

[win4r]

What does this PR do?

Checklist

• Every commit is signed off (git commit -s) — required, CI rejects
  unsigned commits (DCO, see CONTRIBUTING.md)
• Tests pass locally (make -f Makefile.cbm test)
• Lint passes (make -f Makefile.cbm lint-ci)
• New behavior is covered by a test (reproduce-first for bug fixes)

[DeusData]

Thanks @win4r — this is the cumulative branch (it includes #567 + #569 plus the Swift idiomatic type kinds Struct/Enum/Actor, which are a nice addition).

Same blockers as the other two: the README rebrand to a personal fork, the bench/ harness with a hardcoded personal path, the empty description, and DCO not signed.

Rather than three stacked, fork-branded PRs, could you send the useful pieces as focused, signed-off PRs — the explore tool, the Swift fixes/kinds — dropping the README rebrand and bench/? We'd be glad to land those — the engineering is genuinely good; it just needs to arrive clean (no rebrand/bench) and with CI green (signed-off commits, lint and tests passing). 🙏

[DeusData]

Reviewed as the top of the stack — these three are strict containment (#567 ⊂ #569 ⊂ #568), so this PR carries everything in both others. I have closed #567 (fully superseded) and left detailed notes on #569; please read those first. This covers what is unique to #568.

Your extension-clobber catch is sharp, and it may well be a live bug on main.

You spotted that a same-file extension X re-labels struct X back to Class, because the extension pushes a type def with the same qualified name and the UNIQUE(project, qualified_name) upsert makes it last-write-wins. Looking at current main: it still pushes a type def for extension, and since declaration_kind is not struct the label stays Class — so the same clobber mechanism appears to be present. That is a real graph-quality bug and worth its own PR with a repro.

Similarly, main only special-cases Swift struct. enum and actor still lump to Class, so your swift_idiomatic_type_kinds test would be red on main for those two. Genuine repro material.

But the mechanism needs rebuilding before it can land. This PR extends roughly ten hardcoded label allowlists — registry gates in pass_definitions, pass_parallel, pipeline_incremental, both resolve_as_class sites, four store SQL lists, pass_configlink, pass_enrichment, BM25 ranking. Since you branched, main centralised exactly this into cbm_label_is_type_like() (helpers.c:160), whose comment says it plainly: "Single source of truth … adding a label here updates them all." That helper exists precisely to kill the drift that scattered allowlists cause, so re-adding ten of them would move us backwards. A rebase here is a reimplementation, not a conflict resolution.

And one piece is a taxonomy decision rather than a fix. Actor would be a new first-class label in the graph schema — it is not in cbm_label_is_type_like() today. Adding a label has blast radius across every consumer, so that is the maintainer's call and I have escalated it alongside the Swift Enum question. Your instinct matches our stated doctrine (production emits the precise label, consumers get updated rather than the test relaxed) — it is the execution that has aged, not the direction.

Something your work surfaced about main itself, which I am filing regardless of these PRs: store.c still uses label IN ('Function','Method','Class') at four sites (arch boundaries, clusters, vector search). Main already emits Struct nodes for Rust, Go, Swift and D — and those nodes are being excluded from all four today. That is a bug on main that your allowlist audit exposed, and it would have gone unnoticed otherwise. Thank you for that.

Practical asks for any resubmission, all of which currently block CI regardless of merit: 4cd82fa has no Signed-off-by, so DCO fails; the AI session Co-Authored-By trailers should be stripped; and one concern per PR.

One thing that needs independent verification if this concept is revived: the commit itself notes total edges dropping 1813 → 1689 and claims all of them were spurious. That is plausible given the clobber fix, but an edge-count drop is exactly the kind of claim we want measured against a current main rather than inherited.

Leaving this open pending the maintainer's taxonomy answer, rather than closing your work while a decision is outstanding.

[DeusData]

Closing this, and I want to be clear that part of your work has landed on main rather than letting a close read as a rejection.

What shipped, with your credit

#1380 (63eda54, merged as 880b87a6a) carries Co-Authored-By: win4r. Your allowlist audit found that four SQL queries in store.c and the BM25 ranking in mcp.c hardcoded their own label lists, so get_architecture (boundaries, packages, clusters) and vector search were dropping every Struct, Interface, Enum, Type and Trait node in the project — which, for Rust, Go, Swift and D, means most of the codebase's types were missing from the architecture view. That was live on main and nobody else had spotted it.

Two changes from your version, both deliberate. It routes through the canonical cbm_label_is_type_like() set rather than widening the literals in place, and it is pinned by a test in both directions, so the next type-like label fails CI instead of quietly shrinking query results. And it does not include Actor, since main does not emit that label — including it would have been shipping new vocabulary through a bug fix.

Your audit also surfaced a second problem indirectly: while revert-checking that fix, the check passed when it should have failed, because the build did not rebuild on a header change. That is fixed in the same PR.

Why the PRs themselves cannot land

Neither is mergeable in its current shape: both are CONFLICTING against a main that has moved a long way, dco and lint are red, and #568 contains #569 with identical commit OIDs — so merging #568 would silently also take the README rebrand to codebase-memory-mcp-pro and the bench/ harness, which hardcodes a local path and shells out to an unpinned external binary. Those two cannot come upstream in any form.

The nine cherry-picked commits are all already on main under different SHAs, so they are now pure conflict noise.

What is recorded, and still owed to you

Two further bugs you found are real, verified, and on the maintainer's list — they are not being dropped by this close:

1. Swift enum and protocol members are double-emitted. push_class_body_children scans body containers by child type and omits enum_class_body and protocol_body, while find_class_body locates them via the body field — which a child-type scan structurally cannot see. Every Swift enum static and protocol member is emitted twice, once as a Method and once as a spurious top-level Function. Your measurement of 38 duplicate nodes on a 29-file repo checks out, and your fix auto-merges cleanly onto current main.
2. Swift extension clobbers the type it extends. Because Swift qualified names bake in the file stem, struct P {} and extension P {} in P.swift both compute the same QN, and last-write-wins relabels P back to Class and relocates its file_path and line range onto the extension — so get_code_snippet points at the wrong code. Cross-file, extension Date {} mints a phantom Class node for a stdlib type.

The first is held only because it touches the same function as an open architecture decision on another PR — we do not want to answer the same prevention-versus-suppression question twice, in the same function, with different reasoning. When that lands it will carry your credit too.

Left open for the maintainer

The Enum and Actor labelling and the explore tool are direction calls, not defects. Enum is low-risk since it already exists in the canonical set; Actor would be a genuinely new permanent label that every type-resolution consumer would silently drop unless added there as well.

Thank you

Four real defects on main, one of which surfaced only because you audited allowlists nobody else thought to check, and another you reasoned out correctly from the UNIQUE upsert semantics. Your tests pin behaviour rather than asserting it. That is a better hit rate than most review rounds produce, and it is why one of these is already shipped and two more are queued rather than filed away.